When anxiety reaches AI engineers themselves: the fear of AI designing AI

By Zendoric · July 19, 2026. The South Korean newspaper Chosun Ilbo headlines that artificial intelligence researchers feel anxiety about the development of AI systems capable of creating or improving other AI. That is all the available material offers: a headline and a description with no further elaboration, no figures, no names of labs or cited researchers. We are not going to invent what the article does not tell.

What we can put into context, because it is widely documented in the sector, is the underlying phenomenon that headline points to: the automation of AI research itself. Companies like OpenAI, Anthropic and Google DeepMind have long been using their own models to accelerate tasks such as architecture design, hyperparameter tuning, synthetic data generation and debugging of research code. It is not science fiction about an AI that reprograms itself overnight; it is, rather, a progressive automation of technical tasks that previously required PhD students and senior engineers working for weeks.

There lies the irony that probably explains the anxiety reported by Chosun Ilbo: AI engineers themselves, who for years have presented themselves as the group best protected against automation, are beginning to see that their specialty is not immune. If a system can propose and evaluate architecture variants faster than a human team, the researcher's added value shifts toward supervision, the judgment to decide which experiments matter and the interpretation of results, not toward the mechanical execution of the process.

This fits with something we have been pointing out at Zendoric when analyzing the impact of AI on employment sector by sector: the work that gets automated first is the most routine and measurable, and technology work does not escape that logic just by being closer to the tool. In the short term, it is reasonable that there is real concern among junior researchers whose exploration and fine-tuning tasks are the most easily replicated by an automated system.

In the long term, however, if AI truly accelerates its own research, the compound effect could bring forward precisely the advances that underpin our core thesis: better, cheaper models that are faster to develop also mean faster progress in medicine, energy or materials, the areas where abundance is really at stake. The uncomfortable paradox is that the same researchers who feel anxiety about their jobs today could be building, without meaning to, the machinery that shortens the path toward that abundance.